CN105608473A - High-precision land cover classification method based on high-resolution satellite image - Google Patents

Abstract

A land cover classification method of the present invention comprises the steps as follows: (1) determine the patch regions with similar properties of domestic remote sensing images so as to perform reasonable image segmentation; , geometric characteristics, and calculate image features; (3) use historical land use data and feature spectral library, combined with (2) image features, automatically obtain ground feature samples; (4) use (3) ground feature samples as information entropy, Use decision tree and boosting technology to extract various ground features; (5) Calculate the area of various ground features, merge and eliminate the spots according to the set value, and finally obtain the classification result. The invention solves the problems of land cover classification misclassification and many broken spots, improves the operation efficiency by 50%, and improves the classification accuracy to 90%.

Description

A high-precision land cover classification method based on high-resolution satellite imagery

technical field

The invention relates to an object-oriented high-precision land cover classification method based on domestic high-resolution satellite images, which is suitable for land cover classification of multi-spectral images of domestic remote sensing satellites without cloud coverage, and belongs to the field of remote sensing information processing.

Background technique

With the development of the economy, the damage to the global environment is becoming more and more serious. Problems such as droughts and floods, illegal occupation of agricultural land, and deforestation frequently occur and seriously affect people's lives, social and economic development, and social stability. Using remote sensing information to quickly and scientifically , Accurately classify and evaluate the land cover, grasp the situation of regional land cover in a timely manner, and take corresponding countermeasures in time, which play an important role in the rational planning and protection of land resources.

The remote sensing monitoring of land cover is mostly carried out through the monitoring of land cover type, area and change. To know the change of land cover area must be inseparable from the land cover classification. Compared with the extraction of a single type of ground object, remote sensing image There are relatively few algorithms for automatically extracting land cover categories. At present, land cover classification methods such as single-band threshold method, multi-band spectral relationship method, unsupervised classification method, and supervised classification method have been applied one after another.

Among them, the single-band threshold method is a method to distinguish a certain ground feature from other ground features by using the difference in the reflectance or pixel gray value of a certain ground feature and the background feature in a certain band to determine a certain value. The principle of this method is simple and easy to operate. But the key is to determine the threshold, the accuracy of the threshold selection directly determines the accuracy of surface features extraction. Therefore, the single-band threshold method has certain limitations in images with rich types of ground features and close gray values of ground features in the selected band, and its extraction accuracy is low.

The essence of the multi-band spectral relationship method is to construct a band operation function to process the image, and this method can comprehensively extract ground object information by taking advantage of multi-band. This method comprehensively utilizes the spectral information of multiple bands, so the extraction effect is often better than the single-band threshold method. However, this method needs to construct a calculation model of ground feature extraction based on the interspectral relationship method based on the unique multi-band spectral relationship characteristics of ground features in multispectral images of different remote sensing satellites, such as G+R>NIR+MIR or MIR/G<a, etc. , G stands for the green light band, R stands for the red light band, NIR stands for the near-infrared wave band, MIR stands for the short-wave infrared wave band, and a is the threshold. Since the multispectral data of different remote sensing satellites often require different calculation models, this method is not universal and difficult to promote.

The unsupervised classification method is a pixel-oriented classification method, and the operation is relatively simple. However, the classification result of this method is very fragmented, and the amount of data is large, so it is not suitable for high-resolution image classification. Supervised classification generally uses manual sketching of samples, and selects a supervised classification method for pixel-oriented classification. This method requires more manual intervention, more subjective judgments, and the results are more fragmented.

In general, the above methods have the following disadvantages: (1) The method is relatively primitive and the extraction accuracy is low; (2) The extraction process requires auxiliary manual intervention, and even requires on-site measurement and manual sketching, which is time-consuming and laborious ; (3) Different calculation models and thresholds are required for different remote sensing satellite multispectral data, and the universality is low; (4) For high-resolution images, the classification result data of the pixel-oriented method is relatively fragmented, which is not suitable for practical use.

Contents of the invention

The technical problem solved by the present invention is: to overcome the deficiencies of the prior art, and to provide a high-precision land cover classification method based on high-resolution satellite images. The present invention adopts object-oriented thinking to accurately extract the types of ground objects, and can improve production efficiency , the process is simple, and the project is easy to implement.

The technical scheme that the present invention solves is:

A high-precision land cover classification method based on high-resolution satellite imagery includes the following steps:

(1) Determine the patch areas with similar properties in satellite remote sensing images, and perform reasonable image segmentation;

(1a) Calculate the apparent reflectance according to the gray value of the satellite remote sensing image, and the calculation process is as follows:

(1a1) Convert the gray value of the image to the apparent radiance according to the formula:

L _a =Gain×DN+Bias

Among them, Gain is the gain, DN is the gray value, and Bias is the offset;

(1a2) Convert the apparent radiance to apparent reflectance as follows:

${\mathrm{<span\; class="notranslate">\; \&rho;</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; d</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; cos\&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}$

Among them, a represents a positive integer, d is the sun-earth distance correction factor, E _s is the solar spectral irradiance outside the atmosphere, and θ _s is the solar zenith angle;

(1b) According to the spatial resolution of the image, set the image segmentation scale h as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 100</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 30</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 200</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 30</span>}\end{array}\right.$

Among them, h is the image segmentation scale, r is the spatial resolution of the image;

(1c) After determining the image segmentation scale, use the mean shift segmentation algorithm for image block processing to obtain the affected block;

(2) According to the influence block segmented in step (1c), image features are calculated;

The image feature calculation steps are as follows:

(2a) Construct the spectral features of the image features, the spectral features include the average value of the band and the standard deviation of the image spectrum;

Average for each band:

Among them, μ _o is the mean value of the reflectance of the oth band, and o takes a positive integer; ρ _a represents the apparent reflectance of each pixel;

Image spectral standard deviation: ${\mathrm{\&sigma;}}_o=\sqrt{\frac{{\mathrm{\&Sigma;}}_{a=1}^N{({\mathrm{\&rho;}}_a-{\mathrm{\&mu;}}_i)}^2}{N}}$

Among them, σ _o is the image spectrum standard deviation of the oth band;

(2b) Construct the graph features of the image features; the graph features are the geometric characteristics of the image, which are mainly described by the average gradient in this paper;

The average gradient G(x,y) is calculated according to the following formula:

G(x,y)= _dxi + _dyi ;

d _x (i,j)=[ρ(i+1,j)-ρ(i-1,j)]/2

d _y (i,j)=[ρ(i,j-1)-ρ(i,j-1)]/2

Among them, ρ(i+1,j) refers to the apparent reflectance of each band of the image, and i and j represent the rows and columns that affect the apparent reflectance matrix;

(3) Obtain ground object samples according to the image features of step (2);

The sample collection steps are as follows:

(3a) According to the calculated average value, standard deviation and average gradient, collect the whole image in a 3*3 square matrix, and use the following formula to calculate the land type value of the ground object sample pixel:

f(m _i,j )=aμ _i,j +bσ _i,j +cG(x,y)

Among them, f(m _{i, j} ) represents the ground type value of the 3*3 square matrix; μ _{i, j} represents the average value of the band; σ _{i, j} represents the standard deviation of the image spectrum; Influence factor; b indicates the influence factor of standard deviation on feature classification; c indicates the influence factor of average gradient on feature classification; the three coefficients a, b and c are obtained by iterative iteration weighting by using supervised classification method combined with historical feature samples ;

(3b) According to the land type value corresponding to the pixel obtained in step (3a), obtain the surface object sample: based on the actual surface object accuracy, the accuracy of the land type value corresponding to the pixel is within 5% of the actual surface object accuracy Pixels are used as ground object samples;

(4) Use the feature samples in step (3) as information entropy, and use decision tree and boosting technology to extract various features;

(5) Calculating the area of various types of surface feature map spots extracted in step (4), merging and eliminating the map spots according to the set value, and obtaining the final classification result.

The advantage of the present invention compared with prior art is:

(1) The present invention uses the apparent reflectance of the remote sensing image instead of the gray value of the image to extract the object category, because remote sensing imaging detects and collects the radiation of ground objects in the atmosphere, and the atmosphere is the necessary medium for remote sensing information transmission , the solar radiation will have a series of interactions with the atmosphere in the process of atmospheric transmission, which will affect the surface radiance recorded at the entrance pupil of the satellite sensor, that is, the final remote sensing image will deviate from its original surface appearance to a certain extent, Therefore, the atmospheric correction of remote sensing images, that is, the use of apparent reflectance to replace the gray value in the original remote sensing images, can restore the real appearance of the ground surface to a certain extent, and can more accurately extract the object categories in remote sensing images.

(2) The present invention uses an object-oriented method to classify the land cover of the image. In the drifting image segmentation process, combined with the segmentation scale, the image is divided into various parts with the object as the unit based on the characteristics of the image spectrum and texture, so that the classification results of the ground object can be maintained as a whole, and the classification efficiency and accuracy can be improved. The method of the present invention The segmentation scale calculation method takes into account the factor of spatial resolution, which can ensure more accurate and efficient segmentation.

(3) The present invention adopts the process of automatic sample selection and search, which reduces the subjectivity of manual intervention, ensures the accuracy of sample search, and thus achieves the accuracy of classification accuracy.

(4) The invention solves the problems of land cover classification misclassification and many broken spots, so that the operation efficiency is increased by 50%, and the classification accuracy is increased to 90%.

Description of drawings

Fig. 1 is a flow chart of the land cover automatic classification method of the present invention.

detailed description

Specific embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, a high-precision land cover classification method based on high-resolution satellite images includes:

(1) Determine the patch areas with similar properties in satellite remote sensing images, and perform reasonable image segmentation;

(1a) Calculate the apparent reflectance according to the gray value of the satellite remote sensing image, and the calculation process is as follows:

(1a1) Convert the gray value of the image to the apparent radiance according to the formula:

L _a =Gain×DN+Bias

Among them, Gain is the gain, DN is the gray value, and Bias is the offset. These parameters can be obtained from the XML corresponding to the image.

(1a2) Convert the apparent radiance to apparent reflectance as follows:

${\mathrm{<span\; class="notranslate">\; \&rho;</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; d</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; cos\&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}$

Among them, a represents a positive integer, d is the sun-earth distance correction factor, E _s is the solar spectral irradiance outside the atmosphere, and θ _s is the solar zenith angle;

(1b) According to the spatial resolution of the image, set the image segmentation scale h as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 100</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 30</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 200</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 30</span>}\end{array}\right.$

Among them, h is the image segmentation scale, r is the spatial resolution of the image, and the unit is meter. ;

After obtaining the remote sensing image with reflectivity, set the segmentation scale. The setting of the segmentation scale needs to be combined with the spatial resolution of the image. Generally speaking, the image is relatively clear and the ground objects are not broken.

(1c) After determining the image segmentation scale, use the mean shift segmentation algorithm for image block processing to obtain the affected block;

The mean shift has good algorithm convergence, and its direction always points to the place with the maximum local density. At the maximum value of the density function, the drift tends to zero. So the mean shift algorithm is an adaptive fast-ascent algorithm, which can find where the maximum local density is through calculation and "drift" to its position.

If the dimension of the image is p, when the spatial position vector and the color vector are combined into the "space-color" domain, the dimension is p+2, which is expressed as the radial symmetric kernel and the Euclidean multivariate kernel:

${\mathrm{<span\; class="notranslate">\; K</span>}}_{{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; C</span>}}{{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}}\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}{{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; r</span>}}}{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>$

Among them, x ^s is the space part of the feature vector, x ^r is the color part of the feature vector, k(x) uses the same kernel in both space and color domains, h _s and h are the kernel bandwidth and segmentation scale respectively, and C is The corresponding normalization constant. Therefore, the bandwidth parameter (h _s , _hr ) becomes an important parameter in the process of segmentation based on mean shift. And h _s can be obtained through the configuration file corresponding to the image.

The result obtained by this segmentation algorithm divides the image into several blocks, and the obtained image blocks have the characteristics of concentrated objects, good continuity and regular boundaries, which are prepared for the efficiency and accuracy of the following land cover classification.

(2) According to the influence block segmented in step (1c), image features are calculated;

The image feature calculation steps are as follows:

(2a) Construct the spectral features of the image features, the spectral features include the average value of the band and the standard deviation of the image spectrum;

Average for each band:

Among them, μ _o is the mean value of the reflectance of the oth band, and o takes a positive integer; ρ _a represents the apparent reflectance of each pixel;

Image spectral standard deviation: ${\mathrm{\&sigma;}}_o=\sqrt{\frac{{\mathrm{\&Sigma;}}_{a=1}^N{({\mathrm{\&rho;}}_a-{\mathrm{\&mu;}}_i)}^2}{N}}$

Among them, σ _o is the image spectrum standard deviation of the oth band;

The spectral feature can reflect the spectral characteristics of the image, and the overall situation and changes of the image spectrum can be reflected by the average and standard deviation, so that it can be used for sample judgment of land cover classification;

(2b) Construct the graph features of the image features; the graph features are the geometric characteristics of the image, which are mainly described by the average gradient in this paper;

The average gradient G(x,y) is calculated according to the following formula:

G(x,y)= _dxi + _dyi ;

d _x (i,j)=[ρ(i+1,j)-ρ)i-1,j)]/2

d _y (i,j)=[ρ(i,j-1)-ρ(i,j)1)]/2

Among them, ρ(i+1,j) refers to the apparent reflectance of each band of the image, and i and j represent the rows and columns that affect the apparent reflectance matrix;

The average gradient of the image can effectively reflect the spatial details and edge information of the image.

Combining the spectrum, shape and texture characteristics of objects, a series of feature layers with physical meaning and spatial objects as units can be generated to provide rich information support for subsequent object-oriented supervised classification.

(3) Obtain ground object samples according to the image features of step (2);

The sample collection steps are as follows:

(3a) According to the calculated average value, standard deviation and average gradient, collect the whole image in a 3*3 square matrix, and use the following formula to calculate the land type value of the ground object sample pixel:

f(m _i,j )=aμ _i,j +bσ _i,j +cG(x,y)

Among them, f(m _{i, j} ) represents the ground type value of the 3*3 square matrix; μ _{i, j} represents the average value of the band; σ _{i, j} represents the standard deviation of the image spectrum; Influence factor; b indicates the influence factor of standard deviation on feature classification; c indicates the influence factor of average gradient on feature classification; the three coefficients a, b and c are obtained by iterative iteration weighting by using supervised classification method combined with historical feature samples ;

(3b) According to the land type value corresponding to the pixel obtained in step (3a), obtain the surface object sample: based on the actual surface object accuracy, the accuracy of the land type value corresponding to the pixel is within 5% of the actual surface object accuracy The pixel is used as a feature sample; if the feature accuracy of the water body is 1, and the calculated feature value corresponding to the pixel is 0.8, it will exceed 5% of the range and cannot be used as a feature sample of the water body;

(4) Use the feature samples in step (3) as information entropy, and use decision tree and boosting technology to extract various features;

The steps in the land cover classification stage are as follows:

The principle of using the improved SVM (Supportvectormachine) is to use the separating hyperplane as a linear function of separating the training data. SVM allows to directly use the training data to describe the separating hyperplane, which can directly solve the classification problem without density estimation as an intermediate step. Let the training data consist of n samples (l ₁ ,m ₁ ),…,(l _n ,m _n ), l∈Rd, m∈{+1,-1}, separated by the hyperplane decision function:

D(l)=(w·l)+w ₀

Among them, w and w ₀ are decision functions to separate coefficients;

The constraints defining the separability of data samples are:

(w·l _i )+w ₀ ≥+1

If m _i =+1, then:

(w·l _i )+w ₀ ≤-1

If m _i =-1, i=1,...,n, or:

m _i [(w·l _i )+w ₀ ]≥1,i=1,…,n

For a given training data set, the separating hyperplane can be expressed in the above form. The minimum distance from the separating hyperplane to the nearest data point, known as the gap, is denoted by τ. The gap is directly related to the generalization ability of the separating hyperplane. The larger the gap, the greater the separability between classes. Therefore, the condition for selecting the separating hyperplane is to maximize the gap. The support vectors are the data points on the edge of the gap, or equivalently, the data points for which m _i [(w·l _i )+w ₀ ]=1, and the data points closest to the decision surface, which are the most difficult to classify , can determine the position of the decision surface, and the decision surface of the optimal hyperplane can be described by the set of support vectors.

(5) Calculating the area of various types of surface feature map spots extracted in step (4), merging and eliminating the map spots according to the set value, and obtaining the final classification result.

The present invention does not disclose common knowledge in the field.

Claims (1)

1. A high-precision land cover classification method based on high-resolution satellite images, characterized in that it comprises: (1) Determine the patch areas with similar properties in satellite remote sensing images, and perform reasonable image segmentation; (1a) Calculate the apparent reflectance according to the gray value of the satellite remote sensing image, and the calculation process is as follows: (1a1) Convert the gray value of the image to the apparent radiance according to the formula: L _a =Gain×DN+Bias Among them, Gain is the gain, DN is the gray value, and Bias is the offset; (1a2) Convert the apparent radiance to apparent reflectance as follows: ${\mathrm{<span\; class="notranslate">\; \&rho;</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; d</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; \&times;</span>{\mathrm{<span\; class="notranslate">\; cos\&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}}$ Among them, a represents a positive integer, d is the sun-earth distance correction factor, E _s is the solar spectral irradiance outside the atmosphere, and θ _s is the solar zenith angle; (1b) According to the spatial resolution of the image, set the image segmentation scale h as follows: $\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 50</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 100</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 30</span>}<span\; class="notranslate">\; \&rsqb;</span>\\ \mathrm{<span\; class="notranslate">\; 200</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 30</span>}\end{array}\right.$ Among them, h is the image segmentation scale, r is the spatial resolution of the image; (1c) After determining the image segmentation scale, use the mean shift segmentation algorithm for image block processing to obtain the affected block; (2) According to the influence block segmented in step (1c), image features are calculated; The image feature calculation steps are as follows: (2a) Constructing spectral features of image features, said spectral features comprising band averages and image spectral standard deviations; Average for each band: ${\mathrm{\&mu;}}_o=\frac{{\mathrm{\&Sigma;}}_{a=1}^N{\mathrm{\&rho;}}_a}{N}$ Among them, μ _o is the mean value of the reflectance of the oth band, and o takes a positive integer; ρ _a represents the apparent reflectance of each pixel; Image spectral standard deviation: ${\mathrm{\&sigma;}}_o=\sqrt{\frac{{\mathrm{\&Sigma;}}_{a=1}^N{({\mathrm{\&rho;}}_a-{\mathrm{\&mu;}}_i)}^2}{N}}$ Among them, σ _o is the image spectrum standard deviation of the oth band; (2b) Constructing the graph feature of the image feature; the graph feature is the geometric characteristic of the image, which is mainly described by the average gradient in this paper; The average gradient G(x,y) is calculated according to the following formula: G(x,y)= _dxi + _dyi ; d _x (i,j)=[ρ(i+1,j)-ρ(i-1,j)]/2 d _y (i,j)=[ρ(i,j-1)-ρ(i,j-1)]/2 Among them, ρ(i+1,j) refers to the apparent reflectance of each band of the image, and i and j represent the rows and columns that affect the apparent reflectance matrix; (3) Obtain ground object samples according to the image features of step (2); The sample collection steps are as follows: (3a) According to the calculated average value, standard deviation and average gradient, collect the whole image with a 3*3 square matrix, and use the following formula to calculate the land type value of the ground object sample pixel: f(m _i,j )=aμ _i,j +bσ _i,j +cG(x,y) Among them, f(m _{i, j} ) represents the ground type value of the 3*3 square matrix; μ _{i, j} represents the average value of the band; σ _{i, j} represents the standard deviation of the image spectrum; Influence factor; b indicates the influence factor of standard deviation on feature classification; c indicates the influence factor of average gradient on feature classification; the three coefficients a, b and c are obtained by iterative iteration weighting by using supervised classification method combined with historical feature samples ; (3b) According to the land type value corresponding to the pixel obtained in step (3a), obtain the surface object sample: based on the actual surface object accuracy, the accuracy of the land type value corresponding to the pixel is within 5% of the actual surface object accuracy The pixel is used as a feature sample; if the feature accuracy of the water body is 1, and the calculated feature value corresponding to the pixel is 0.8, it will exceed 5% of the range and cannot be used as a feature sample of the water body; (4) Use the feature samples in step (3) as information entropy, and use decision tree and boosting technology to extract various features; (5) Calculating the area of various types of surface feature map spots extracted in step (4), merging and eliminating the map spots according to the set value, and obtaining the final classification result.